Instructions to use water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir MiniMax-H3-MLX-Argus-Calibrated-INT8 water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
| { | |
| "schema_version": 3, | |
| "profile": "m4-pro-calibrated-int8", | |
| "algorithm": { | |
| "family": "activation-aware diagonal-Hessian affine weight-only INT8", | |
| "objective": "minimize the diagonal-Hessian approximation to ||WX-Q(W)X||_F^2", | |
| "optimizer": "three alternating weighted least-squares affine refits on Torch/CUDA", | |
| "fallback": "retain the plain MLX RTN group whenever a fitted group is not better", | |
| "packing": "four consecutive uint8 codes per uint32, first code in least-significant bits" | |
| }, | |
| "bits": 8, | |
| "group_size": 32, | |
| "mode": "affine", | |
| "quantize_adaln": true, | |
| "adaln_bits": 8, | |
| "quantized_layers": { | |
| "8": 254 | |
| }, | |
| "bf16_layers": [ | |
| "blocks.14.adaln_proj.linear", | |
| "blocks.46.mlp.fc2", | |
| "blocks.47.mlp.fc2", | |
| "blocks.48.mlp.fc2" | |
| ], | |
| "bf16_exceptions": [ | |
| { | |
| "path": "blocks.47.mlp.fc2", | |
| "parameters": 77070336, | |
| "rtn_diagonal_hessian_relative_error": 0.006842044550077006, | |
| "full_dit_parameter_fraction": 0.002326792637428219, | |
| "selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget" | |
| }, | |
| { | |
| "path": "blocks.14.adaln_proj.linear", | |
| "parameters": 260112384, | |
| "rtn_diagonal_hessian_relative_error": 0.006829738475327525, | |
| "full_dit_parameter_fraction": 0.00785292515132024, | |
| "selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget" | |
| }, | |
| { | |
| "path": "blocks.46.mlp.fc2", | |
| "parameters": 77070336, | |
| "rtn_diagonal_hessian_relative_error": 0.006813439715784915, | |
| "full_dit_parameter_fraction": 0.002326792637428219, | |
| "selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget" | |
| }, | |
| { | |
| "path": "blocks.48.mlp.fc2", | |
| "parameters": 77070336, | |
| "rtn_diagonal_hessian_relative_error": 0.006765240564342134, | |
| "full_dit_parameter_fraction": 0.002326792637428219, | |
| "selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget" | |
| } | |
| ], | |
| "bf16_exception_parameters": 491323392, | |
| "bf16_exception_full_dit_parameter_fraction": 0.014833303063604896, | |
| "protected_source_precision_modules": [ | |
| "video_patch_proj", | |
| "audio_patch_proj", | |
| "condition_proj", | |
| "time_embedder", | |
| "final_layer", | |
| "norms", | |
| "linear_biases" | |
| ], | |
| "calibration": { | |
| "path": "/root/.argus-skill-latest/workspaces/s-4e0f9076/research/minimax-h3-int8-calibration-cuda-fp32-r939557dc.npz", | |
| "sha256": "0ff2892e605979bd7a2fe0c2807bdd05d9c2b1c54af1bcd53bd3f00b046d18dc", | |
| "manifest": { | |
| "algorithm": "diagonal-hessian-activation-aware-affine-int8", | |
| "bits": 8, | |
| "cases": [ | |
| { | |
| "height": 256, | |
| "prompt": "A red fox leaps over a mossy log in a misty forest at dawn.", | |
| "seed": 101, | |
| "split": "calibration", | |
| "step_index": 0, | |
| "width": 256 | |
| }, | |
| { | |
| "height": 192, | |
| "prompt": "Waves crash against black volcanic rocks under a grey sky.", | |
| "seed": 202, | |
| "split": "calibration", | |
| "step_index": 2, | |
| "width": 320 | |
| }, | |
| { | |
| "height": 320, | |
| "prompt": "A street musician plays saxophone on a rainy neon-lit corner.", | |
| "seed": 303, | |
| "split": "calibration", | |
| "step_index": 4, | |
| "width": 192 | |
| }, | |
| { | |
| "height": 192, | |
| "prompt": "Steam rises from a bowl of noodles on a wooden table.", | |
| "seed": 404, | |
| "split": "holdout", | |
| "step_index": 1, | |
| "width": 256 | |
| }, | |
| { | |
| "height": 256, | |
| "prompt": "An origami robot walks across a sunlit library desk.", | |
| "seed": 505, | |
| "split": "holdout", | |
| "step_index": 3, | |
| "width": 192 | |
| } | |
| ], | |
| "collector": { | |
| "cuda": "13.1", | |
| "cuda_visible_devices": "1", | |
| "dit_load": { | |
| "assigned_parameter_tensors": 638, | |
| "compute_dtype": "float32", | |
| "dropped_source_keys": [ | |
| "rope.inv_freq" | |
| ], | |
| "load_seconds": 39.6426787157543, | |
| "missing_keys": [], | |
| "peak_cuda_memory_bytes": 132720563200, | |
| "source_tensors": 535, | |
| "unexpected_keys": [] | |
| }, | |
| "framework": "PyTorch/CUDA FP32 DiT teacher with one-time BF16 conditioner", | |
| "gpu": "NVIDIA B200", | |
| "script": "scripts/calibrate_int8_torch.py", | |
| "text_encoder": { | |
| "encode_seconds": 1.0854376470670104, | |
| "parity": { | |
| "comparison": null, | |
| "passed": null, | |
| "performed": false, | |
| "prompts": {}, | |
| "tolerance": { | |
| "cosine": 0.997, | |
| "relative_l2": 0.08 | |
| } | |
| }, | |
| "peak_cuda_memory_bytes": 50327620608, | |
| "token_ids_exact_between_source_assets": true, | |
| "torch_loader": { | |
| "full_layers": 64, | |
| "load_seconds": 10.279082759749144, | |
| "loaded_layers": 50, | |
| "loaded_parameter_tensors": 552, | |
| "missing_keys": [], | |
| "unexpected_keys": [] | |
| } | |
| }, | |
| "torch": "2.11.0a0+eb65b36914.nv26.02", | |
| "transformer_mapping": "official convert_transformer_key + reorder_interleaved_qkv; PyTorch hooks mapped back to original MLX module paths" | |
| }, | |
| "created_at": "2026-08-11T16:27:54.492972+00:00", | |
| "group_size": 32, | |
| "implementation": { | |
| "commit": "db1d7651847c2b382beae2094d28214712ff811e", | |
| "dirty": true, | |
| "repository": "https://github.com/Argus-AiTeam/minimax-h3-mac" | |
| }, | |
| "mode": "affine", | |
| "sampling": { | |
| "adaln_selection": "real video/audio scheduler timesteps at preregistered case indices, recorded separately from the full per-forward timestep table", | |
| "adaln_split_timesteps": { | |
| "calibration": [ | |
| 0.0, | |
| 0.05263161659240723, | |
| 0.1818181872367859, | |
| 0.25, | |
| 0.5714285373687744 | |
| ], | |
| "holdout": [ | |
| 0.020408153533935547, | |
| 0.07692301273345947, | |
| 0.1111111044883728, | |
| 0.3333333134651184 | |
| ] | |
| }, | |
| "duration_seconds": 0.2, | |
| "max_rows_per_call": 8, | |
| "max_rows_per_layer": 64, | |
| "selection": "evenly spaced rows from each real layer input", | |
| "steps": 6 | |
| }, | |
| "schema_version": 1, | |
| "source": { | |
| "model": "MiniMaxAI/MiniMax-H3", | |
| "revision": "939557dc319dd91227e30195a763f272ba7f8765", | |
| "transformer_config_sha256": "f619093a231fcfbcc3d035bec26c50ad864e7331a500d5c519f5045dc1e50458" | |
| }, | |
| "split_contract": "calibration and holdout prompts, seeds, and recorded AdaLN timestep inputs are disjoint", | |
| "token_ids": { | |
| "A red fox leaps over a mossy log in a misty forest at dawn.": [ | |
| 32, | |
| 2518, | |
| 38835, | |
| 83458, | |
| 916, | |
| 264, | |
| 77243, | |
| 88, | |
| 1487, | |
| 304, | |
| 264, | |
| 8600, | |
| 88, | |
| 13638, | |
| 518, | |
| 38393, | |
| 13 | |
| ], | |
| "A street musician plays saxophone on a rainy neon-lit corner.": [ | |
| 32, | |
| 8592, | |
| 38744, | |
| 11088, | |
| 63008, | |
| 77203, | |
| 389, | |
| 264, | |
| 62757, | |
| 46652, | |
| 2852, | |
| 275, | |
| 9131, | |
| 13 | |
| ], | |
| "An origami robot walks across a sunlit library desk.": [ | |
| 2082, | |
| 2713, | |
| 10606, | |
| 12305, | |
| 22479, | |
| 3941, | |
| 264, | |
| 7015, | |
| 31635, | |
| 6733, | |
| 18010, | |
| 13 | |
| ], | |
| "Steam rises from a bowl of noodles on a wooden table.": [ | |
| 41644, | |
| 37168, | |
| 504, | |
| 264, | |
| 19212, | |
| 315, | |
| 60160, | |
| 389, | |
| 264, | |
| 22360, | |
| 1965, | |
| 13 | |
| ], | |
| "Waves crash against black volcanic rocks under a grey sky.": [ | |
| 54, | |
| 4693, | |
| 9920, | |
| 2348, | |
| 3691, | |
| 66064, | |
| 23035, | |
| 1212, | |
| 264, | |
| 19780, | |
| 12884, | |
| 13 | |
| ] | |
| } | |
| } | |
| }, | |
| "calibration_objective": { | |
| "plain_rtn_diagonal_hessian_sse": 4572725.832609705, | |
| "candidate_diagonal_hessian_sse": 2366819.81746003, | |
| "relative_improvement": 0.4824050458959487 | |
| }, | |
| "optimizer": { | |
| "framework": "Torch/CUDA", | |
| "device": "cuda:0", | |
| "gpu": "NVIDIA B200", | |
| "torch": "2.11.0a0+eb65b36914.nv26.02", | |
| "cuda": "13.1", | |
| "calls": 512, | |
| "seconds": 958.1897342111915, | |
| "peak_memory_bytes": 16988678144, | |
| "layer_residency": "one dense source layer and its FP32 calibration rows are uploaded at a time; CUDA tensors are released after MLX uint32/scales/biases are materialized" | |
| }, | |
| "source": { | |
| "model": "MiniMaxAI/MiniMax-H3", | |
| "revision": "939557dc319dd91227e30195a763f272ba7f8765", | |
| "subfolder": "FL2VA/transformer", | |
| "config_sha256": "f619093a231fcfbcc3d035bec26c50ad864e7331a500d5c519f5045dc1e50458" | |
| }, | |
| "builder": { | |
| "repository": "https://github.com/Argus-AiTeam/minimax-h3-mac", | |
| "repository_commit": "db1d7651847c2b382beae2094d28214712ff811e", | |
| "repository_dirty": true, | |
| "mlx_version": "0.32.0", | |
| "python": "3.12.3", | |
| "platform": "Linux-6.8.0-1043-nvidia-x86_64-with-glibc2.39", | |
| "implementation_sha256": { | |
| "scripts/calibrate_int8.py": "d29123d6982475574b01bca1ceffa15770f507ea44d6bdb1779bb53ce745faa2", | |
| "scripts/calibrate_int8_torch.py": "693a2b12af17ccc75124db649334af7f1dff6d51afec47ed6b9aacd0207c9892", | |
| "scripts/probe_real_cuda_parity.py": "5798dcb3fed3e69ceed8f0157fa6bb7ce8ffad8eda4590947e5819b431919ad3", | |
| "scripts/activation_quant_torch.py": "ab7ec7def56b5cb442743732dd16e9d2d74ebe3228261e85c72c427c2adcc64a", | |
| "scripts/build_calibrated_int8.py": "49108bf521b40e11477d4e6f908a877810d75487990b5cad54b063a1c362ff70", | |
| "scripts/eval_calibrated_int8.py": "81205559dd460dfaa4eb09eac796907b1fb4f7cafb70a43009f8391472ce5bf1", | |
| "scripts/build_quant.py": "a0d1a552286f30e3219596bd5f06e5b99fb0eeefdf3b952202152f3ffcdac106", | |
| "minimax_h3_mlx/activation_quant.py": "37621b33250fa2cb497e5cd71fc63bd124a050336d27de9e3b4154df9bc77b10", | |
| "minimax_h3_mlx/quantize.py": "c908638eeec83c127a1b90f93aef607e5d12563f78011702a78465a2010ba275", | |
| "minimax_h3_mlx/load.py": "4a87519b5795f92c08e5f8daf31e023b28f6491a32744a48afe804a211cbc92d", | |
| "minimax_h3_mlx/streaming.py": "dbf76e338bd4c4f9a01b2f8517039b74a0f0953a3f20bebd4ae39eac43fef2c3", | |
| "minimax_h3_mlx/dit.py": "a75791c95244ae496490f7ae0275e3ac42007506bec7b467ad74c56ca8f55c46" | |
| } | |
| }, | |
| "gb_on_disk": 37.797, | |
| "gb_resident_after_adaln_drop": 22.928 | |
| } |